Cosmo (Amazon's Rufus-Powering Algorithm)

Cosmo is Amazon's algorithm powering amazon-rufus (its conversational AI shopping assistant), described as distinct from the legacy keyword-intent-focused Amazon A9 algorithm rather than a replacement for it. Sellers still optimizing purely for A9-style keyword density (see Keyword Ecosystem Validation and Master Keyword List & Listing Scorecard) are said to be leaving conversational-search visibility on the table.

Mechanism

Dynamically routes each Rufus query to a different underlying LLM via Amazon Bedrock (e.g., Amazon Nova, Anthropic models) depending on whether the query is a simple lookup or research-heavy.
Layers Retrieval-Augmented Generation (RAG) on top of a base LLM, pulling live external data rather than relying solely on trained knowledge.
Draws on five data sources: catalog data, reviews, Q&A, external web/publication sources (curated review/publication sites — notably not Reddit or Quora), and customer behavior, reportedly weighted toward roughly the last 2-3 weeks rather than all-time history.

Knowledge-graph mapping

Cosmo is said to require listings to carry a technical attribute-node mapping (a 'knowledge graph') — e.g., a product tagged as 'used with' complementary items or 'used in location X' — so it surfaces across conversational queries that never mention its literal keywords. This reframes keyword-stuffing as a secondary signal beneath structured attribute mapping, while leaving room for more brand-story-driven listing copy aimed at human conversion.

Provenance note

Public detail on Cosmo currently comes largely from vendor-interview content (e.g., ZonGuru) promoting paid optimization services, rather than from Amazon's own technical documentation — treat specific figures (e.g., $10B incremental sales, 60% purchase-likelihood lift) as vendor-sourced claims pending independent confirmation.

Rufus Image & Review Scanning

Rufus (the shopper-facing conversational AI) scans product images for keywords/attributes and mines customer reviews for sentiment when answering shopper questions, on top of parsing listing text — meaning image content and review sentiment now function as de facto SEO inputs, not just conversion aids. Apply: listing images should carry keyword-bearing benefit callouts (ingredient → benefit) and sellers should actively manage review sentiment, since both are now machine-read by Cosmo/Rufus rather than only viewed by humans.

Из тем: Unsorted, SEO & Keyword Strategy: Winning A9, Cosmo & Rufus